Classification
A radar detection by itself tells you something is there. What it is is what matters for the driving decision — pedestrian, cyclist, car, lorry, motorbike. Classification is Provizio’s deep-learning module that applies object-class labels to every detection in the radar point cloud, combining proprietary DSP with convolutional neural networks trained on diverse radar datasets.
300+
objects classified per frame
<25 ms
added latency
6
object classes
Applications
Where the perception layer needs to label what it detects.
Automotive
ADAS and autonomous driving — classify vehicles, pedestrians, cyclists, motorbikes and large vehicles. Class labels drive everything from AEB target selection to lane-change risk assessment.
Industrial
Mining, agriculture, construction — separating pedestrian workers from large machinery in close proximity is a safety-critical classification task. Provizio’s Industrial-trained classifier handles this where automotive-only models fail.
All-weather
Radar perception works through fog, smoke, dust, glare and darkness. Classification operates wherever the radar does — when cameras and LiDAR are degraded, class labels stay available.
Infrastructure
Roadside and smart-intersection perception. Classify vehicles, pedestrians and cyclists at junctions and along the road for traffic monitoring and roadside safety — in any conditions, day or night.
How it works
How the radar point cloud becomes labelled objects.
STEP 01
Read the radar point cloud
Each radar frame arrives as a 3D point cloud — thousands of returns, each carrying its position, how fast it is moving from Doppler, and how strong its signal is. This is the raw material the classifier works from.
STEP 02
Turn it into a signature image
Provizio translates the point cloud into a bird’s-eye image, where each object becomes a distinctive colour signature — its motion and signal strength painted straight into the picture, ready for recognition.
STEP 03
Recognise every object
A trained neural network reads the signature image and names what it sees — large vehicles, cars, pedestrians, motorbikes and cyclists — learning from diverse radar data through the URMAP flywheel.
STEP 04
Label every object, frame by frame
Each object carries a class label and a confidence score, published over DDS at the radar’s frame rate — ready for the tracking, path-planning and collision-avoidance systems downstream.
Featured deployment
A pedestrian, classified in total darkness.
In this capture the camera sees almost nothing — yet radar perception still picks out and classifies a pedestrian crossing in the dark. Because radar does not depend on light, class labels stay available through darkness, glare, fog, rain and dust — exactly where camera-based classifiers drop out. It builds a strong foundation for the night-time demands of FMVSS 127 pedestrian AEB and beyond.
- Pedestrian classified in near-total darkness
- Same scene, camera effectively blind
- Robust through darkness, glare, fog, rain and dust
- A strong foundation for FMVSS 127 night pedestrian-AEB
Integration
Runs on GPU or SoC, ships via DDS
Classification is a microservice in the 5D Perception stack, running on the available compute — at the edge or in centralised compute, depending on your architecture. For constrained edge deployments, the URMAP Lite variant is optimised for the low-power compute on a radar SoC. Each object’s class and confidence are published over DDS at the radar’s frame rate, alongside the detection, tracking, odometry and freespace streams.
- Runs on GPU or SoC — edge or central
- DDS class-label output
- Optimised for real-time inference
- Trained via URMAP data flywheel (PaaS)
- OTA model updates for new classes
- Inherits URMAP variants (Auto / Industrial)
Questions, answered
Classification FAQ
The baseline class set covers large vehicles, cars, pedestrians, motorbikes and cyclists. Custom classes can be added through Perception as a Service — Provizio fine-tunes the URMAP classifier on customer-provided data for application-specific class sets (e.g. industrial categories like haul trucks, dumpers, dozers).
Classification runs on a GPU, or an embedded SoC for the URMAP Lite edge variant. Whether it sits at the edge — on or beside the radar — or in centralised compute depends on your system architecture. For edge deployments, our URMAP Lite variant provides a version of Classification optimised for the low-power compute you’d typically find on a radar SoC. Processing latency scales with the compute you run it on.
Radar maintains performance through fog, smoke, dust, heavy rain, snow, darkness and direct sunlight. Classification operates whenever the radar produces a point cloud, which means class labels are available across conditions where camera-based classifiers fail.
Same CNN architecture; different trained weights. The Automotive variant is tuned for road users (vehicles, pedestrians, cyclists). The Industrial variant is trained on construction, mining and agricultural object classes — including the safety-critical task of separating workers from large metal machinery at close proximity.
Yes — through Perception as a Service. As part of PaaS, the URMAP data flywheel uploads field data automatically; Provizio labels it and fine-tunes the model on top of URMAP’s universal base for your application-specific classes, then deploys the improved model back to your radar over the air.
Detect, classify, decide — on the edge.
Classification gives your perception stack radar-native, object-level class labels that stay reliable through weather, darkness and at range — where camera and LiDAR degrade. Let’s talk about integrating it into your pipeline.